The best AI coding agents are rarely the most capable ones

I read a number this week that would make most engineering budgets flinch. On Vercel’s platform, an open-weight model most executives have never heard of, Z.ai’s GLM-5.2, grew about 27 times in daily token volume and about 80 times in customer count in its first full week. Not over a quarter. A week.

That is not a story about one model. It is a story about how enterprises actually decide what a coding agent runs on, and it quietly contradicts the thing most leadership teams believe about picking the best AI coding agents.

The belief that the best coding AI agents are the most capable ones

Here is the assumption I hear in almost every tool conversation. The company picks the smartest model, puts every engineer on it, and treats capability as the thing it is paying for. The board approves the premium tier because nobody wants to explain later why the company shipped on the second-best coding agent.

It feels responsible. It is also, increasingly, not how the best coding AI agents get chosen inside serious engineering orgs.

Why “buy the best AI coding model” feels like the safe call

The instinct is understandable. For two years, capability really was the constraint. The gap between the top model and everything else was wide enough that paying up for the best AI coding model was the rational move. The premium tier was buying the difference between “usable” and “not.”

Nobody ever got questioned for buying the model at the top of the leaderboard. They just got a bill.

The problem is that the ground moved and the instinct did not. The capability spread between the leaders has compressed to the point where, for most everyday coding tasks, three or four models all clear the bar. When several options all do the job, the smartest one stops being the obvious answer and starts being the expensive one.

What the routing data says about the best AI agents for coding

This is where the week got specific. CNBC reported that Chinese open-weight models now account for as much as 46 percent of the tokens US firms route through the OpenRouter gateway, holding above 30 percent every week since February, up from an 11 percent average before that. These are not hobby projects. That is enterprise coding work moving to models that cost a fraction of the frontier tier.

The reason is not politics. It is arithmetic. Leading open-weight models are running 60 to 90 percent cheaper than the top US systems, some as low as 18 cents per million tokens against roughly 4 dollars for a frontier model. And GLM-5.2 landed within a single percentage point of Claude Opus 4.8 on a closely watched agentic benchmark at about a fifth of the cost. On SWE-bench Pro it posted 62.1 percent, ahead of GPT-5.5 at 58.6.

Harpreet Arora, who runs agentic infrastructure at Vercel, put the mechanism plainly to CNBC: “Price is doing the work here. When a task doesn’t need the best model, teams are beginning to route it to the cheapest one that’s good enough.” The recent wave of open models coming out of China, he added, is winning that trade.

The same logic shows up inside the premium vendors too. This week Anthropic’s most capable coding model, Fable 5, went back on the meter at a rate where a single agentic session that generates 2 million output tokens costs about 100 dollars in credits. The same session on Sonnet 5 runs closer to 20. The reporting that walked through those numbers called Sonnet 5, not the flagship, the economically rational default for most enterprise coding work. When the vendor’s own math points away from the top of its own lineup, “best” has clearly stopped meaning “most capable.”

And the payoff does not keep climbing the way the premium framing implies. A developer survey published back in April found the productivity curve flattens within a couple of months of adoption, whichever tool sits underneath.

If most of the gain lands early and then plateaus, the ongoing premium for the top-tier model is buying a lot less than the leaderboard suggests.

Where “best AI for coding, free” quietly became an enterprise strategy

The part that should reframe the whole conversation is that a chunk of this shift runs on open weights. “Best AI for coding, free” used to be a search a student typed. Now it is a line in a platform strategy, because open weights give a self-host option that cuts the data path and the per-token bill at the same time. That is not a China story and it is not a discount story. The same routing logic applies to every model on the menu, US or not, hosted or local.

The leaderboard is the least durable thing an org can standardize on. The routing policy is the asset.

Best now means cheapest-good-enough plus a name on the verification

So here is the reframe I would offer any executive staring at a tool decision. This is not really about adopting a model. It is about building two disciplines, and the model is a swappable part inside them.

The first is routing: a policy that decides which class of task goes to which tier of model. Cheap and good-enough for the bulk work, frontier only where the task genuinely needs it. The second is verification: a named human who owns whether the output is correct, because a cheaper model that occasionally slips only saves money if someone catches the slip before it merges.

That is genuinely good news for the budget, by the way. The winning move here is not exotic. It is just good management with a clearer picture of the bill.

Three moves before you standardize on one agent

None of this requires a reorg. It requires three small decisions before the next renewal.

First, split the coding work into “needs the frontier model” and “does not,” honestly. Most orgs are surprised how much lands in the second bucket. That split is the routing policy in one page.

Second, put a real number on what the premium tier is actually buying. If the top model and a model at a fifth of the cost both clear the same verification bar on a class of task, the premium is reassurance, not results. Name that trade out loud.

Third, put one human’s name against verification for each service area. The whole cost advantage of routing to a cheaper model depends on catching its mistakes, and a dashboard does not catch anything.

The best AI coding agents for your team, it turns out, are not sitting at the top of anyone’s leaderboard. They are the ones a routing policy sends the right work to, checked by someone whose name you can say. That is a calmer decision than the one the benchmarks push for, and it holds up a lot longer.